DT-Autoencoder

Usage

Four possible input arguments:

1. --training_data: arff file with training data, check Input Data section for more info.
2. --testing_data: arff file with testing data, check Input Data section for more info.
3. --outfile: json file that stores results with trained and initial weights.
4. --rand_init: use this flag to use random weight initialization.

Use --rand_init flag to start with random parameters, Recall and AUC otherwise.

Example:

The following command will train the model using the breast-cancer.training1.arff file and test using breast-cancer.testing1.arff. The results will be saved in my_result.json. The weights will be randomly initialized.

python main.py --training_data breast-cancer.training1.arff --testing_data breast-cancer.testing1.arff --outfile my_result.json --rand_init

Dependencies

1. python=3.6.7
2. python-weka-wrapper3
3. Anaconda

Input Data

Only for categorical variables.

Training dataset should contain Class Attribute last, should be only one class (e.g. genuine).

Testing dataset should contain Class attribute last, should be binary labels (e.g. impostor, genuine but any name will work as long as it is binary). impostor Label should be first.

Changing the classifier

The Classifier for the encoder is a J48 Decision Tree implementation from Weka. You can edit the arguments by editing the options list or change the classifier altogether on line 28 of classifier.py within the train_trees() function definition. For information on the available options for J48 you can look here. Right now the default values <pruning confidence> = 0.25 and <minimum number of instances> = 2 have been set.

For information on other decision tree implementations within Weka you can check here. Using any of these should be as simple as changing the classname and option arguments as necessary.